A Hybrid UKF with Explicit Drivetrain Modeling for Low-Velocity State Estimation of Electric Vehicles

summary

The gist

Longitudinal velocity is estimated using an unscented Kalman filter on a double-track model adapted to low velocities by incorporating explicit drivetrain modeling and data-driven virtual sensors to…

In short

This work estimates electric vehicle longitudinal velocity at low speeds using an unscented Kalman filter (UKF) on a double-track model. It improves accuracy by incorporating explicit drivetrain modeling and data-driven virtual sensors, overcoming the limitations of standard wheel-speed sensors. The resulting estimators show significant error reduction, demonstrating robust cross-vehicle transfer capabilities.

Key concepts

Double-Track Model
A mathematical framework used to estimate vehicle state by combining two different models. In this paper, one track uses explicit drivetrain modeling (motor speed as a state) and the other uses a neural network to simulate wheel speeds, allowing for better estimation at low velocities.
Explicit Drivetrain Modeling
This involves adding detailed physics of the electric motor and transmission directly into the estimation model. This makes the motor speed an estimated state rather than just an input, which is crucial for accurately determining vehicle velocity near standstill where wheel sensors fail.
Unscented Kalman Filter (UKF)
A sophisticated statistical tool used to estimate a system's state when measurements are noisy. The UKF handles non-linear systems better than simpler filters, making it ideal for using complex models like the double-track model and incorporating both physical dynamics and data from virtual sensors.
Zero-Shot Transfer
The ability of the trained estimators (Elastic or GRU) to work immediately on a new vehicle without needing retraining. This proves the robustness of the estimation method, as they maintain their learned parameters even when deployed on a different vehicle with different dynamics.

Terminology used across episodes

This episode discusses

The paper

A Hybrid UKF with Explicit Drivetrain Modeling for Low-Velocity State Estimation of Electric Vehicles · Read on arXiv

Institute of Mechatronic Systems, Leibniz University Hannover · IAV GmbH

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "A Hybrid UKF with Explicit Drivetrain Modeling for Low-Velocity State Estimation of Electric Vehicles".

Dev: Here's the gist of the paper. Longitudinal velocity is estimated using an unscented Kalman filter on a double-track model adapted to low velocities by incorporating explicit drivetrain modeling and data-driven virtual sensors to overcome limitations of production wheel-speed sensors Estimation Strategy The work proposes estimating longitudinal velocity with an unscented Kalman filter (UKF) on a double-track model adapted to low velocities by incorporating the motor speed along two routes, one purely physical and one hybrid The first route augments the model with an explicit torsional spring-damper drivetrain, making the motor speed a state of the estimator The second route feeds the output of a neural network-based virtual wheel-speed sensor into the UKF as a pseudo-measurement Both routes are compared against two baselines, a wheel-speed-only configuration and one that adds the motor-speed sensor but treats the driveline as rigid Model and Discretization The estimator rests on a planar double-track model with Magic Formula tire forces and rotational wheel dynamics This structure is extended by an explicit drivetrain model that couples the electric machine to the driven axle Near standstill, any such slip-based model turns numerically stiff The discretization study benchmarks fixed-step schemes against an adaptive-step reference solution to select a configuration that is stable down to standstill Comparison of Estimators The four configurations are compared against the conventional wheelspeed-only baseline (Wheels) and against the baseline that adds the motor-speed sensor to Wheels with the simplest coupling, a rigid driveline without torsion (Rigid) The comparison shows that both proposed estimators lower the velocity root-mean-square error by at least 54 % and 45 % against both baselines on Vehicle A below 1 km/h Cross-Vehicle Validation Both estimators carry across vehicles without a new identification campaign because they are tuned on Vehicle A and deployed on a second vehicle, Vehicle B, without re-tuning The GRU virtual wheel-speed sensor leads in every maneuver class below 1 km/h and lowers the pooled error in this target regime by 52 % when transferred to Vehicle B The work demonstrates that both proposed estimators carry across vehicles without a new identification campaign Elastic is preferable when the driveline parameters of the target vehicle are accessible and the estimate must remain interpretable GRU is preferable when those parameters are uncertain but for which recordings from a comparable platform exist Both proposals reach a similar accuracy in the target regime, so the choice between them rests on practical differences The low-velocity model, discretized by a scheme chosen in a systematic comparison of explicit and semi-implicit integrators, stays stable down to standstill at real-time cost The deployed configuration of this work is the explicit Euler scheme with n = 100 sub-steps Tuning and Noise The process noise Q is tuned per variant on the training recordings defined in Section V-B using a particle-swarm optimization (PSO) The tuning integrates with the same explicit Euler scheme at n = 100 sub-steps that is later deployed The tuned noise is static, adapting neither over time nor with speed Experimental Setup Both test vehicles are series-production battery-electric vehicles driven by a single rear-mounted motor through a fixed single-speed transmission The inertial reference is an Automotive Dynamic Motion Analyzer (ADMA-G-ECO+) sampled at 50 Hz The evaluation is restricted to three maneuver classes: low-speed maneuvers, cornering at medium speed, and curb crossings The reported quantity is the root-mean-square error (RMSE) of the longitudinal velocity against the ADMA reference Results Summary In the sub-1 km/h target regime, Elastic reduces RMSE(vx) by 55 % from Rigid to 0.044 m/s GRU reduces it by 46 % to 0.052 m/s The curb class separates the variants most clearly, with Elastic reaching 0.051 m/s against 0.085 m/s for Rigid GRU reaches the lowest pooled RMSE(vx) of the four variants at 0.038 m/s in the cornering class The zero-shot transfer to Vehicle B shows that GRU leads in every maneuver class below 1 km/h, at 0.029 m/s for low speed Elastic is second in all three classes on the transfer vehicle Limitations Two limitations bound these results: first, the tuned noise is static, which leaves a speed-scheduled or online-adapted process noise as the natural next step Second, the virtual sensor enters the filter as a point estimate, so its measurement noise is fixed a priori The diagonal R ignores the correlated errors of both rear wheel channels The stability floor for the explicit torsion model is found at n = 40 sub-steps, and the deployed n = 100 sits directly at this limit The explicit torsion state therefore costs about 2.5 times the sub-steps of the rigid model Appendix Details The closed-loop Euler step-size sweep shows that the rigid-driveline baseline is stable from n = 40 sub-steps, whereas the elastic driveline requires n = 100 for every run to converge The diverging runs occur only on the curb-crossing maneuvers, whose load changes excite the drivetrain torsion The sweep confirms the Euler-100 deployment anchor of Section III in closed loop and quantifies the cost of the explicit torsion model The deployed n = 100 therefore sits directly at the stability limit and gains nothing from finer stepping The stability floor is a property of the stiff dynamics, not of the tuning Acknowledgment The authors used Anthropic’s Claude in the preparation of this article All content was reviewed and verified by the authors The two vehicle illustrations in Fig.

Title and authors: Dev: Let's talk about the paper "A Hybrid UKF with Explicit Drivetrain Modeling for Low-Velocity State Estimation of Electric Vehicles." They are focusing on how to get accurate speed estimates when you're operating in that low-speed regime where conventional sensors fail.

Rosa: The title itself points to the solution: combining a hybrid unscented Kalman filter with an explicit model of the drivetrain. It’s about making sure you have a better handle on what's happening inside the vehicle when things are slow and subtle >

Taro: So, is this just about making the math look fancier, or is there something tangible that changes how a safety system actually works in those low-speed situations?

Dev: It’s about making the motor speed a state in the filter because it stays informative down to standstill, even when the wheels are reporting noise >

Rosa: They are essentially building this dual estimation system—the physical route with torsion and the data-driven route with a neural network sensor—and then comparing them against simpler setups to see which one actually performs better on real vehicle A >

The paper's summary: Rosa: So, if we boil down what they did, they took a double-track model and adapted it for low speeds. They built two estimation routes: one route adds an explicit torsional spring-damper drivetrain to make the motor speed a state, and the second route uses a gated recurrent unit neural network as a virtual wheel speed sensor input >

Dev: And they put both of those routes head-to-head against two baselines. The first baseline is just using wheelspeed, which fails below one kilometer per hour, and the second baseline adds the motor speed sensor but assumes the driveline is rigid with no torsion at all >

Taro: That comparison structure seems important because it isolates whether the improvement comes from modeling that physical compliance or from learning something new about what the wheels are actually doing >

Rosa: Exactly. They showed that both of these proposed estimators manage to lower the root-mean-square error by at least fifty four percent and forty five percent against both of those baselines when testing on Vehicle A below one kilometer per hour >

The paper's improvements: Dev: The key improvement they highlight is that both of their estimators can carry over their tuning—the process noise, the drivetrain parameters, and the trained network—to a second vehicle without any new identification campaign >

Rosa: That cross-vehicle transfer is what makes this really interesting for deployment. They tuned everything on Vehicle A and then deployed it on Vehicle B without re-tuning >

Taro: So they claim this means you can test something once, tune it, and then have a reasonably good system ready for another vehicle without spending weeks re-identifying the dynamics >

Dev: That's the point. They also selected their integration scheme through a systematic discretization study, choosing an explicit Euler scheme with one hundred sub-steps because it stayed stable down to standstill at real-time cost >

Conclusion: Rosa: So to wrap up, the paper "A Hybrid UKF with Explicit Drivetrain Modeling for Low-Velocity State Estimation of Electric Vehicles" shows that both the explicit torsional drivetrain route and the GRU virtual sensor route work well, though they have different strengths depending on your situation >

Dev: They found that if you know the driveline parameters well, like in Vehicle A, the Elastic model is preferable because it keeps things interpretable >

Taro: But if those physical parameters are uncertain or you don't have recordings from a comparable platform, then the GRU route is better because it learns something useful and can transfer across vehicles >

Rosa: Both proposals reach similar accuracy in that low-speed regime, so the choice really comes down to practical differences between knowing your vehicle's physics versus relying on learned data >

Dev: And they confirmed that the low-velocity model, discretized by that specific scheme chosen in their comparison, stays stable down to standstill at real-time cost using a deployment of one hundred sub-steps with an explicit Euler scheme >

Taro: So for us, it means we have two paths: use the physics if you can measure it precisely, or rely on the learned component if you need something that can just work across different setups >

More episodes

← Home